About parameter issues

[Songchuanlang]

Thank you very much for the code library you provided. Now I have two questions when reproducing the fsood detection.

• There are some problems with the training and testing parameters of your sem method. I saw that some topics mentioned that the parameters of mnist are set to -1, 0, 1, and the settings of cifar are -0.01, 0, 1. However, when using the cifar dataset as the ID dataset, the ResNet-18 five-layer network is used, so your parameter settings are -0.0001, 0, 0, 0, 1, and whether the bias added when calculating the score can be changed to adjust the score. So when I use pca50 for dimensionality reduction, I can't reach the AUROC in the paper. Setting it to pca10 can exceed the AUROC in the paper.

• There is no test script for fs-ood in the training script and test script, which is difficult for me to reproduce. I hope to get some tips and help from you.
  

I don't know if this is correct.

[Jingkang50]

Thank you for your interest in our work.

• Yes, this method could be sensitive to the hyperparameter, therefore setting number as -1e-5 can make improvement in cifar
• yes, fs-ood use gmm.yml to test

[Songchuanlang]

Thanks for your reply.What is the difference between these two calculations, or when to use them?
in GMM.postprocessors

    # conf = torch.matmul(score_list, alpha)
    conf = torch.matmul(torch.log(score_list + 1e-45), alpha)

[Songchuanlang]

I have some other questions, can you answer them together? Why do we need to train the Gaussian mixture model again when calculating the score?